NHB 2026 · 论文互动解读Interactive paper walkthrough
AI 生成 · 示意图，非实验照片 / AI-generated illustration
Nature Human Behaviour · 2026 · Nielsen & Christiansen

你脑子里存的，
不只是“语法树”
Your brain stores more than grammar trees

6 项研究、900 多人的数据显示：像 “added to a” 这样不成短语的三个词，也会在大脑里留下“走熟的小路”。这是不是说明——人也像大模型一样“预测下一个词”？今天我们一起判断。
Six studies with more than 900 people show that even three-word strings that are not phrases, like “added to a”, leave well-trodden paths in the mind. Does that mean humans, like LLMs, just “predict the next word”? Let’s decide together.

← → 或滚动翻页 · 按 S 显示讲者提示 · 右上角切换语言Arrow keys or scroll · press S for speaker notes · switch language top right

开场：先念网络上流行的那句话——“原来人类也是根据上文，预测下一个词”。告诉大家今天要做的事：用这篇论文检验这句话，结论在倒数第三页揭晓。森林小路是作者自己在论文里用的比喻，后面会讲。
先投票Quick poll

“原来人类也是根据上文，预测下一个词。”你同意吗？
“Turns out humans also just predict the next word from context.” Agree?

按举手人数逐次点击；结果只存在本机浏览器，最后一页再投一次做对比。Click once per raised hand; counts stay in this browser. We vote again at the end.

别评价任何一方，让大家带着自己的选择往下听。
背景：语言学的老共识Background: the old consensus

传统观点：句子像一棵树，词先组成“短语”（成分），再层层拼成句子
The classic view: a sentence is a tree — words form phrases (constituents) that nest into a sentence

点一点：下面哪几组词是树上完整的一根树枝（成分）？Tap: which word groups are a whole branch of the tree (a constituent)?

点上面的词组试试。Tap a word group above.
像 “boy gave the”“book to the” 这种半截的、跨了树枝的词串，叫“非成分”。传统理论里，它们不是大脑存储的单位。
Strings like “boy gave the” or “book to the” are cut across branches: non-constituents. In classic theories they are not units the mind stores.

示意据论文图 1a 的例句重绘（非原图）。Schematic redrawn from the example sentence in the paper’s Fig. 1a (not the original figure).

给大众：成分就像“整块乐高”，非成分像把两块乐高各掰半块拼一起。给医生：生成语法（乔姆斯基）和构式语法都基本以成分为中心，这篇对两派都提出挑战。
论文要测的东西What the paper tests

不看具体的词，只看“词性的排队顺序”
Ignore the words; look only at the order of parts of speech

addedVERB 动词toPREP 介词aDET 冠词
≈
definedVERB 动词byPREP 介词theDET 冠词

词完全不同，意思也无关，但“动词-介词-冠词”这串排队一样。Different words, unrelated meaning, same VERB–PREP–DET line-up.

核心问题

如果大脑里真的存着这种线性的词性小片段，那么刚看过一个，再看另一个同样排队的词串时，应该会快一点。这叫“结构启动”。

关键：这些三词串几乎都不是短语——在语料库里，实验 1、2 的词串平均只有 1.1% 和 0.9% 的时候以“成分”出现。

The question

If the mind really stores such linear part-of-speech chunks, then after seeing one, a different string with the same line-up should be processed a little faster. That is “structural priming”.

Key point: these strings are almost never phrases — in the corpus, the Study 1 and 2 strings occurred as constituents on average only 1.1% and 0.9% of the time.
给医生：启动效应和临床里的“启动”是一回事吗？For clinicians: is this the same “priming” as in clinical tests?
同一思路：先前的加工痕迹让后续相同类型的加工更快、更省力。语义启动（看到“医生”后“护士”认得更快）测的是意义；这里测的是结构，所以作者特意让启动词和目标词没有重复的词，并在语义相似度、长度、词频上做了匹配（实验 4 只按长度匹配）。
Same logic: an earlier processing trace makes later processing of the same kind faster. Semantic priming (“doctor” speeds “nurse”) tests meaning; this tests structure, so primes and targets share no words and were matched on semantic similarity, length and frequencies (Study 4 on length only).
一句话记住：同样的“词性队形”会互相加速。
亲自试一下：实验任务Try the task yourself

三个词，能不能出现在一个正常英语句子里？越快越好
Three words: could they be part of an English sentence? Answer fast

—
点“开始”。每个词串之间有 500 毫秒的“+”。Press Start. A “+” shows for 500 ms between strings.
真实实验：每人做 320 次判断，其中藏着 64 对“启动-目标”。比如先看 knew that she（动词-连词-代词），紧接着看 wondered if you（同样队形）；对照组则先看 example of an（另一种队形）。
The real study: 320 judgements per person, with 64 hidden prime–target pairs. E.g. knew that she (verb–conjunction–pronoun) right before wondered if you (same line-up); controls saw example of an first.

⚠ 课堂体验，不是论文数据。几十毫秒的差别，一个人做几次是测不出来的。⚠ Classroom demo, not study data. A few tens of milliseconds can’t be detected from a few trials.

请一位观众上台按键。重点不是测出效应，而是让大家感受“任务就是判断词串是否像英语”，真实实验里参与者是点按钮作答，F/J 键只是课堂演示。
实验 1–2：效应存在，而且看“顺序”Studies 1–2: the effect exists, and order matters

同样队形快了约 33 毫秒；把队形倒过来，加速就没了
Same line-up: about 33 ms faster. Reverse the order and the boost is gone

实验 1（40 人）Study 1 (n = 40)

启动效应 33.0 毫秒（95% 可信区间 3.9–61.1），证据比 ER = 71.3。Priming effect 33.0 ms (95% CrI 3.9–61.1), evidence ratio 71.3.

实验 2（57 人）：加入“倒序”组Study 2 (n = 57): adds a reversed order

同序比对照快 39.2 毫秒（区间 −10.5 到 88.6，ER = 16.0）；比倒序快 63.8 毫秒（8.4 到 119.2）。例：me again and（代词-副词-连词）→ or how you（连词-副词-代词）。词性一样，顺序相反。Same vs control 39.2 ms (CrI −10.5 to 88.6, ER 16.0); same vs reversed 63.8 ms (8.4 to 119.2). E.g. me again and (PRON–ADV–CONJ) → or how you (CONJ–ADV–PRON): same parts, reversed.

倒序组反而比对照组慢了 24.3 毫秒（证据只算弱到中等）。所以加速不是因为“又见到了动词、介词”，而是因为同样的先后顺序。
The reversed order was even 24.3 ms slower than control (weak-to-moderate evidence). So the boost is not from seeing the same word classes again, but from the same sequence.

数据：论文 Study 1–2 正文（后验中位数）。坐标轴不从 0 开始，以便看清几十毫秒的差别。Data: Study 1–2 text (posterior medians). Axis does not start at zero, to show differences of tens of ms.

“顺序很重要”这一点，是后面联系大模型时的伏笔：LLM 也是按顺序读文本的。但别急着下结论。
实验 3（200 人）：成分 vs 非成分Study 3 (n = 200): constituents vs non-constituents

“不成短语”的词串，启动效应一点不弱，甚至更强
Non-phrase strings prime just as well — if anything, more strongly

36.5毫秒 · 非成分
20.5毫秒 · 成分

非成分更强的证据为中等（ER = 6.2）。也就是说，“是不是完整短语”不是这种加速的前提。

36.5ms · non-constituent
20.5ms · constituent

Moderate evidence that non-constituent priming is stronger (ER = 6.2). In other words, being a complete phrase is not required for the boost.

未启动Not primed非成分 · 已启动Non-constituent primed成分 · 已启动Constituent primed

数据：Study 3 正文。注意两组反应时基线不同，比较的是各自的“加速量”。Data: Study 3 text. Baselines differ; compare each group’s speed-up.

这一页回答的是“会不会其实是短语在起作用”。答案：不是。
实验 4（200 人）：大脑会不会“自动补全”？Study 4 (n = 200): does the mind auto-complete?

看到 “live in a”，你会不会先在心里补上 “car”？
Seeing “live in a”, do you silently fill in “car”?

live in a ?

点“?”看看：这正是“预测下一个词”的直觉。如果大脑总在补全，那“半截”(part) 和“完整短语”(whole) 在心里应该是一回事。Tap “?”: that’s the “predict the next word” intuition. If the mind always completes, the part and the whole should be the same thing inside.

part（半截）：live in a（非成分）
whole（整体）：live in a car（完整短语）
part: live in a (non-constituent)
whole: live in a car (full phrase)
结果：半截目标被“半截”启动时，比被“整体”启动时快 27.6 毫秒（95% 可信区间 −1.6 到 57.2，略跨 0；ER = 32.5）；整体目标反过来也一样（27.6 毫秒，区间 −1.6 到 56.7，ER = 29.8）。方向可信度高，但大小不确定。作者据此认为：这与“先预测、补全成整个短语”的机制不一致——半截本身就是被单独存储的。
Result: part targets were 27.6 ms faster after part primes than after whole primes (95% CrI −1.6 to 57.2, just crossing 0; ER 32.5), and vice versa for wholes (27.6 ms, −1.6 to 56.7; ER 29.8). The direction is credible; the size is uncertain. The authors conclude this is inconsistent with a structural-prediction mechanism that completes the phrase — the part is represented on its own.
给医生：可信区间跨过 0，为什么还说“证据强”？For clinicians: the interval crosses 0 — why “strong evidence”?
论文用的是贝叶斯混合模型。ER（证据比）= 效应方向为正 vs 为负的后验概率之比；ER > 10 被作者标为强证据。95% 区间轻微跨 0（如 part→whole 24.8，区间 −7.0 到 57.6，ER = 14.9）意味着方向较可信，但大小不确定。读的时候请同时看点估计、区间和 ER，别只看星号。
Bayesian mixed models. ER = posterior odds that the effect is in the stated direction; the authors flag ER > 10 as strong. An interval slightly crossing 0 (e.g. part→whole 24.8, −7.0 to 57.6, ER 14.9) means the direction is fairly credible but the size uncertain. Read estimate, interval and ER together.

数据：Study 4 正文与补充表 1。Data: Study 4 text and Supplementary Table 1.

这是全课最重要的一页，和那句流行语直接相关：作者专门检验了“补全/预测”这种解释，结果不支持。但注意，它否定的是“预测整个短语来解释这种启动”，不是说人从不预测。
走出实验室Out of the lab

真实阅读和真实打电话里，同样的痕迹也在
The same trace shows up in real reading and real phone calls

眼动阅读（68 人，读新闻句子）Eye-tracked reading (n = 68, news sentences)

上一句里出现过一个相同的“非成分队形”，下一句的总阅读时间少约 48 毫秒（区间 −15 到 147，跨 0；ER = 17.1，方向较可信），回看次数也少（ER = 35.8）。但“第一遍阅读时间”没有明确变化（ER = 4.1）。
One shared non-constituent line-up in the previous sentence cut total reading time by about 48 ms (CrI −15 to 147, crossing 0; ER 17.1: direction fairly credible) and reduced regressions (ER 35.8). First-pass reading time showed no clear change (ER 4.1).

电话对话（358 人，642 通电话）Phone conversations (358 speakers, 642 calls)

某个非成分队形上一轮刚出现过，下一轮说出同样队形的概率 2.12% → 2.39%，比值比 OR = 1.13。After a non-constituent line-up appeared in the previous turn, the chance of it being uttered next turn rose 2.12% → 2.39%, OR 1.13.

而且“非成分”能启动“成分”（OR = 1.29），反之亦然（OR = 1.19，区间 0.96–1.45，证据较弱）：同一串词性，不管有没有组成短语，都会被“传染”。
Non-constituents also primed constituents (OR 1.29) and vice versa (OR 1.19, CrI 0.96–1.45, weaker evidence): the same line-up spreads whether or not it forms a phrase.

数据：Study 5–6 正文。语料分析是观察性的“准实验”，已排除逐字重复的情况。Data: Studies 5–6. Corpus analyses are observational quasi-experiments; verbatim repeats were excluded.

给大众的翻译：聊天时我们会不自觉地“跟着对方的句式走”，而且跟随的单位比短语还碎。
作者的解释The authors’ explanation

语言加工像在森林里走路：走得越多，小路越好走
Processing language is like walking through a forest: the more a path is used, the easier it gets

added to a defined by the VERB · PREP · DET
作者的设想：每读或说一句话，大脑都在“意义空间”里走一条轨迹。两句话词性排队相同，轨迹就部分重合；重合的那段被反复使用，就像被很多动物踩出的小路，越来越好走。刚走过的路，再走就更快——这就是启动。

旁证：越常见的词性队形，判断越快（不论有没有启动）。
The authors propose: every utterance traces a path through “meaning space”. Two utterances with the same part-of-speech line-up share part of the path; that shared stretch, used again and again, becomes a well-trodden trail — easier each time. A recently walked trail is faster to walk again: that’s priming.

Supporting hint: more frequent line-ups were judged faster, regardless of priming.
“加工胶水”（作者的推测）：“on the top” + “top of the” + “the tallest mountain” → “on the top of the tallest mountain”。半截词串像胶水，把两个完整的意思块粘在一起。
“Processing glue” (authors’ speculation): “on the top” + “top of the” + “the tallest mountain” → “on the top of the tallest mountain”. The fragment glues two meaningful chunks together.

“轨迹/小路”是作者提出的解释，森林比喻出自讨论部分；“胶水”出自讨论部分及图 7（此处为重绘示意），作者明确称其为推测。The path account is the authors’ proposal, with the forest analogy from the Discussion; the “glue” idea is from the Discussion and Fig. 7 (redrawn schematic) and is explicitly speculative.

强调：这一页是作者的“解释和推测”，不是被实验直接证明的机制。
回到开头那句话Back to the claim

“原来人类也是根据上文预测下一个词”——这篇论文能证明吗？
“Humans also predict the next word from context” — does this paper show it?

先猜，再点卡片翻面。Guess first, then tap each card.

我的结论（讲者判断）：一半对。论文支持“人脑存着线性的词性小片段，且越常见越快”，作者认为这类片段可能经由统计学习获得；但论文没有测“预测下一个词”。它检验的“先预测、补全成短语”只是被排除为这种启动的解释。作者同时承认，人本来就擅长结构预测。所以“人会预测”没被否定，“人就是在预测下一个词”也没被证明。My verdict (presenter’s judgement): half right. The paper supports that the mind holds linear part-of-speech chunks, processed faster when more frequent, which the authors think could be acquired by statistical learning. But it does not test next-word prediction. The prediction-completion account it tested was ruled out only as the explanation of this priming, and the authors note people are proficient at structural prediction. So “humans predict” is not refuted, and “humans just predict the next word” is not shown.
讲者观点 · 非论文结论Presenter’s view · not the paper’s
这种“从经验里学到词怎么排队”的思路，和大模型的学习方式有相似之处。更准确的说法也许是：“人和大模型都会从大量语言经验里学到‘词怎么排队’的统计规律。”至于人在多大程度上靠“预测下一个词”来理解语言，是另一批研究的问题；而且行为上相似，不等于机制相同——人说话还有意图、身体、情境和社交目的。
This “learn how words line up from experience” idea does resemble how LLMs learn. A fairer slogan: “Both people and LLMs pick up statistical regularities of how words line up from lots of language.” How much humans rely on next-word prediction is a question for other research; and similar behaviour is not the same mechanism — people speak with intentions, bodies, context and social goals.
翻卡顺序建议：1→2→3→4→5→6。第 3 张最容易引起争论：作者在引言里提到，循环神经网络和现代大模型能复现很多结构启动现象，但作者也说“还不清楚这些模型用的是什么表征”。
局限 · 给医生 · 再投一次Limits · for clinicians · vote again

别夸大它Don’t over-sell it

  1. 只研究了英语母语者；作者推测能推广到其他语言，但需要验证。Only English native speakers; generalisation is expected but untested.
  2. 只看三四个词的片段；作者承认，更长的整句里“成分”可能仍然重要。Only 3–4-word strings; constituents may still matter for full sentences, the authors note.
  3. 效应只有二三十到几十毫秒，是群体平均，不是个人能感觉到的。Effects are tens of milliseconds, group averages, not something individuals notice.
  4. 实验 4 的材料只能按长度匹配（补充分析显示结果稳健）；分析没有设盲；排除标准未预注册。Study 4 matched only on length (robust in supplementary checks); analysis not blinded; exclusion criteria not preregistered.
  5. “森林小路”“加工胶水”是解释和推测，作者引用了 Chang 等人的神经网络模拟作为旁证，但没有脑成像或本研究自己的模型验证。“Forest paths” and “glue” are interpretations, supported indirectly by Chang et al.’s network simulations, but not tested with brain imaging or models in this study.
给医生 · 讲者引申，非论文结论For clinicians · presenter’s extension
临床上，部分失语患者仍能说出“套话”和高频词串（如问候语、口头禅），这与“多词片段被单独存储”的思路相容。不过套话是固定的词串，而本文研究的是抽象的词性队形。这篇论文没有研究患者，但它提醒我们：评估语言时，流利的套话不一定代表组句能力完好。
Clinically, some people with aphasia still produce formulaic phrases and high-frequency strings. That fits the idea of separately stored multiword chunks — though stock phrases are fixed word strings, while this paper concerns abstract part-of-speech sequences. This paper studied no patients, but it is a reminder that fluent stock phrases may not mean intact sentence building.

听完了，再投一次Vote again

“原来人类也是根据上文，预测下一个词。”“Humans also just predict the next word from context.”

课前Before

课后After

讨论题Discussion

你说话时，有没有发现自己总在用某几句“顺嘴”的句式？那是你在“预测”，还是在“走熟路”？Do you notice “easy” sentence patterns you keep using? Is that predicting, or walking a familiar trail?

比较课前课后投票。理想的变化是从“同意/不同意”的两极，走向“一半对”。
来源与说明Sources and notes

论文与图片来源
Paper and image credits

Nielsen, Y. A. & Christiansen, M. H. Evidence for the representation of non-hierarchical structures in language. Nature Human Behaviour 10, 579–588 (2026). https://doi.org/10.1038/s41562-025-02387-z

配套简报：Research Briefing: Priming of non-constituents reveals linguistic structure beyond grammar. Nat. Hum. Behav. 10, 457–458 (2026). https://doi.org/10.1038/s41562-025-02388-y

数据与代码：Data and code: https://osf.io/ymndq/

图是怎么来的
· 封面森林：AI 生成的示意图，非实验照片，不承载数据。森林小路是作者在论文中使用的比喻。
· 句法树、词性卡片、小路动画、“胶水”示意：按论文例句用代码重绘，不是原图。
· 所有数据图：用代码根据论文正文和补充材料中的数字绘制（后验中位数及 95% 可信区间）。无逐人数据，所以没有画散点。
· 论文为订阅文章（非 CC BY），本课件未复制任何原图。
· 虚线框内容为讲者引申，非论文结论。
How the visuals were made
· Cover forest: AI-generated illustration, not a photo, carries no data. The forest-path analogy is the authors’.
· Syntax tree, POS cards, path animation, “glue” schematic: redrawn in code from the paper’s examples, not original figures.
· All data charts: drawn in code from numbers in the main text and Supplementary Information (posterior medians, 95% CrIs). No per-person data, so no scatter points.
· The article is subscription content (not CC BY); no original figures were copied.
· Dashed boxes are the presenter’s extensions, not the paper’s conclusions.
← 返回学习中心← Learning center